Remote Sensing Image Dehazing Based on Dual Attention Parallelism and Frequency Domain Selection Network

Abstract

Remote sensing (RS) image dehazing holds immense importance for enhancing the utility of remote sensing technology across both military and civilian domains. Due to the ineffective utilization of multi-scale and frequency information, image dehazing often struggles to handle the uneven distribution of haze in remote sensing images. To address this problem, a dual attention parallelism and frequency domain selection network (DAFSNet) is proposed in this paper. The DAFSNet consists of two primary components, namely the Dual Attention Parallel (DAP) module and the Frequency Domain Selection (FDS) module. The DAP module leverages channel attention and multi-scale pixel attention mechanisms to extract both globally shared information and multi-scale local spatial details associated with haze-related features. Meanwhile, the FDS module decomposes the extracted features into independent frequency information and dynamically selects useful frequency components through the fusion attention mechanism. These two modules are integrated to capture both multi-scale spatial domain features and efficient frequency domain features of RS images, thereby facilitating the efficacious restoration of haze-free images. Experimental results on SateHaze1k and RICE datasets prove that the proposed DAFSNet outperforms the existing state-of-the-art (SOTA) methods.